Vision-based target detection for AGV docking

Abstract

Nowadays, with the development of science and technology, there are more and more application scenarios for Automatic Guided Vehicles (AGV), which has an irreplaceable position in the development of intelligent manufacturing. A widely used scenario is that AGV replaces manpower to perform pickup/delivery tasks in factory floors and warehouses. AGVs are required to navigate accurately and perform precise docking operations. In the docking process, positioning and target identification are vital technologies. Many methods have been developed to solve this problem, such as line patrol, UWB, SLAM, etc. This dissertation aims to design a perpendicular docking system with an accuracy of ~2cm. Pure 2D LIDAR docking technology has the shortcoming of low target recognition accuracy because the point cloud scanned by LIDAR can only tell the shape information of the objects. When the objects are similar in shape, it will cause false detection. To improve the accuracy of target detection, vision-based object detection is used to assist in identification. The camera can obtain more semantic information, which can well enhance the stability of detection. By combining the detection results of the camera with the point cloud of LIDAR, the docking system can be more robust. Vision-based object detection has traditional algorithms (such as SIFT, ORB) and learning-based methods. Learning-based object detection outperforms the traditional ones in accuracy. However, considering AGV’s low-end GPU, standard CNN models can’t operate on such devices for their huge computing requirement. Hence, lightweight networks will be alternative to perform real-time object detection with lower computational cost. In this project, a lightweight model, SSD-MobileNetV3 will be designed and trained according to the specific objects inside our lab environment. Then the model will be applied to the AGV to test the perpendicular docking. The final result achieves an accuracy of ~1cm.